Chinese Journal of Computers   Full Text
Title	Combo-Dimensional Kernels for Graph Classification
Authors	LI Yu-Feng1) Kwok　James Tin-Yau2) ZHOU Zhi-Hua1)
Address	1)(National Key Laboratory for Novel Software Technology, Nanjing University, Nanjing 210093) 2)(Department of Computer Science and Engineering, Hong Kong University of Science & Technology, Hong Kong)
Year	2009
Issue	No.5(946—952)
Abstract &   Background	Abstract Learning from structured data, such as graphs, is an important problem in machine learning. Kernel method is regarded as a powerful solution to such a problem. This paper focuses on molecular graph classification and, following Swamidass et al.’s work, proposes an improved method using combo-dimensional kernels. The proposed method first constructs 2D kernels combined with 1D information to describe chemical characteristics, and to describe physical characteristics, it then constructs 3D kernels based on geometrical information and related molecular mechanics knowledge. Furthermore, inspired by ensemble learning with multiple dimensions, the method finds the optimal kernel combination by quadratically constrained quadratic programming. Experiments show that the proposed method outperforms existing algorithms. Keywords machine learning; graph classification; kernel methods; structure information; ensemble learning Background Sign language recognition aims to translate sign language to text or speech, so as to bridge the communication between the deaf and the hearing people and help the deaf or hard-of-hearing better integrate into the society. According to data collection of sign language, sign language recognition is generally divided into two major categories: Dataglove-based sign language recognition and vision-based sign language recognition. Vision-based sign language recognition attracts more attention of the researchers for it is more convenient for users than Dataglove-based sign language recognition. However, the state of art of vision-based sign language recognition is far from real application due to the difficulty of extracting efficient sign language features from video or image. One of great challenges in vision-based sign language recognition is to achieve view independent sign language recognition because view variance brings feature variance. Most of the current methods in vision-based sign language recognition require a specific view of the signers, generally the frontal view, so as to guarantee similar features between the training samples and the test samples. The constraint of a specific view means that the signers can only perform with specific location and orientation, and limits the freedom of the signer. The authors aim to achieve viewpoint independent sign language recognition within a certain scope with only one camera, so as to remove the restriction of the specific view and provide convenience for users. In the previous works, the authors based on the epiplor geometry and proposed the efficient method of verifying the uniqueness of fundamental matrices for viewpoint free sign language recognition. However, the method of verifying the uniqueness requires at least 8 point correspondences between one frame pair and will fail under data deficiency. So the authors propose the Sample-Consensus method in this paper for efficient viewpoint free sign language recognition under data deficiency. This research is sponsored by the National Natural Science Foundation of China under contract Nos.60533030, 60603023 and U0835005 and partially sponsored by open project of Beijing Multimedia and Intelligent Software Key laboratory in Beijing University of Technology. One purpose of these projects is to solve the key problems in sign language recognition.